As of October 2026, ChatGPT runs two different shopping systems. In the US, products come from feeds that OpenAI indexes itself. In Europe, they still come from Google Shopping. That means brands need different shopping optimization strategies for each market.
We found out about the discrepancy through a ChatGPT debug API that returns around 150,000 lines of retrieval data per prompt, including the score ChatGPT gives each product. It revealed the full product retrieval journey in both regions. Malte Landwehr, CPO and CMO at Peec AI, presented the findings at brightonSEO 2026. Watch the talk below.
TL;DR
Two geographies mean two systems. US shopping runs on OpenAI’s own index, Labrador, fed by product feeds. OpenAI still scrapes Google Shopping in Europe.
Unverified sellers are cut in the US. Under OpenAI’s Labrador index, a shop without the verified-seller flag drops out of results completely.
Descriptions stop at 50 tokens. Anything past roughly the first 40 words of a product description is ignored at ChatGPT’s scoring stage.
Brands should send every product variant to OpenAI. In the US feed-based system, each model and color can take its own slot.
Europe uses Google Shopping for now. Using a third party means no proprietary relevance scoring and no seller check, so in Europe, brands’ performance on Google Shopping feeds straight into ChatGPT.
A quick glossary. The shopping carousel is the block of product cards in a ChatGPT answer (it includes one to seven products, sometimes in table format). Merchants are the shops you can buy from when you click a product. Sources are the web pages ChatGPT uses to write the text around the carousel.
The US and Europe now get different shopping results
A few weeks before Malte’s talk in Brighton, OpenAI switched US shopping to its own retrieval system. The debug data shows that US sessions have a flag for Labrador shopping version 2 that is set to true, and a SERP API shopping scraping flag set to false. European sessions show the reverse of these flags.
US | Europe | |
|---|---|---|
Product source | OpenAI's Labrador index, built from product feeds | Google Shopping, scraped via a SERP API |
How brands get in | Send a feed to OpenAI or sell on Shopify | Be listed in Google Shopping |
Seller check | Verified-seller flag required | None |
Relevance scoring | Embeddings, BM25, GTIN matching | None. Results assigned round robin across fan-out queries |
Time to retrieve | About 0.6 seconds | About 4 seconds |
Click tracking |
| No feed parameter |
The practical upshot is that if you do well in ChatGPT Shopping in Europe today, that may simply mean you do well in Google Shopping. When Labrador reaches Europe, that advantage likely will not persist.
In the US, your product feed decides whether you show up
Where does ChatGPT Shopping get its products in the US? At the time of the talk, US shopping drew on 669 feeds, 606 of them active. About 70 came from direct partnerships, where a brand like Nike sends its feed to OpenAI. Around 500 were still feeds OpenAI built itself by scraping retailer websites through Bright Data. The rest came through Shopify, plus one from Salesforce and one from a Stripe integration. When a brand signs a direct partnership with OpenAI, OpenAI switches off the feed it had built by scraping that brand’s website.
How does ChatGPT pick products? We ran a test prompt for running shoes. ChatGPT pulled 41 candidate products, turned 32 of them into detailed prompt blocks (with title, description, ratings, global trade item number [GTIN], and more), and showed 5 products as results. For each candidate, the debug data shows embedding scores, a BM25 score, approximate nearest neighbor (ANN) retrieval, and reciprocal rank fusion (RRF). These are standard information retrieval methods, and you can run BM25 on your own product data locally to get a rough sense of how well your listing scores.
Unverified sellers vanish from results. Each candidate product carries an ```is_site_verified_seller attribute. Without it, a shop drops out of the results completely. One shop in our test had 25 of the 32 candidates but placed zero products in the final results, because it had no verified seller flag. We haven’t yet found what determines the flag’s status.
More listings can mean more carousel slots. In our test, Dick’s Sporting Goods had five candidate products, and four made the final five. Nike had two, and one made it. De-duplication only removed the same shoe in different sizes. That points to a different strategy from SEO, where the habit is one landing page per product. In a feed-based system, it pays to send every item and variant as its own entry. More entries mean more chances to take carousel space from competitors.
Product descriptions stop at 50 tokens. Every product description in a prompt block is cut at exactly 50 tokens. One from our test ended mid-sentence: “Pattern offers a smooth heel to…” In English, 50 tokens equals roughly 40 words or 250 characters, so put your most important terms first.
Personalization is built in. The shopping system knows the user’s location (both home and current location), whether the precise location is shared, an estimated gender with a confidence score, and which merchants and marketplaces the user has already seen or clicked. Results get more personal over time.
Amazon is being tested. The debug data lists the A/B tests OpenAI is running, and one is called “No Amazon”: US shopping with zero Amazon results. At this stage, it’s only a test, but it shows how quickly the merchant mix could change.
Sources come from several indexes. Besides picking products, ChatGPT also searched the web to write the text around the carousel. In our test, that search turned up 42 candidate sources, and ChatGPT used one of them (Runner’s World). Labrador has separate indexes for the web, Reddit, news, PDFs, YouTube, and Wikipedia. Pulling from an index doesn’t mean ChatGPT prefers it in citations. Often it’s just context for writing the answer.
Older, smaller models do the retrieval. ChatGPT’s user-facing answer uses the flagship model, but product retrieval, selection, and prompt-block writing run on custom versions of GPT-4o, GPT-4.1 nano, and GPT-5 mini in a single session.
In Europe, Google Shopping still feeds ChatGPT
Europe’s pipeline is simpler and slower. For one test prompt, ChatGPT sent five fan-out queries (the background searches an AI platform runs to answer one prompt) to Google Shopping through a SERP API. They returned between 1 and 40 products each.
ChatGPT then merged the results of the fan-out queries so that it took the first product from each query, then the second from each query that still had a product remaining, and so on, until it had a list of 113 products. In its answer, it showed 18. The whole step took about four seconds.
Relying on Google Shopping means there are no embeddings, BM25, or GTIN matching applied to results, and no verified-seller check. Any store in Google Shopping counts as trustworthy for ChatGPT in Europe. So for now, ChatGPT performance in the region is identical with Google Shopping performance. But that won’t always be the case, and we think OpenAI will eventually make the switch to its own index. Running BM25 on your product data is worth doing early, before the change happens.
What to do now
Get your feed to OpenAI. You can send it directly, or sell through Shopify. Scraped feeds are a stopgap, which OpenAI is working to replace with partnerships.
Put key info in the first 40 words of product descriptions. ChatGPT retrieves very little of these texts, and will cut them off at a hard 50 tokens.
List every product variant separately. Each color or model is another chance at a slot in the shopping carousel.
Run BM25 on your product data. It’s free, runs locally, and shows how relevant your listings score with a key piece of ChatGPT Shopping’s retrieval system. This is easy to do before Labrador reaches Europe.
Keep your Google Shopping presence strong in Europe. Until Labrador arrives, it remains the source of ChatGPT Shopping’s European results.
Tag feed clicks. Look for
utm_medium=feedin your analytics to attribute US shopping traffic.
Slides from the talk
How to track which products appear in ChatGPT Shopping
Track the prompts your customers ask, then measure how often and where each product appears in the answers. Peec AI’s AI Shopping Analytics does this at the product level: you upload your catalog, and Peec AI matches it against the chats it tracks every day. Visibility shows the share of chats in which a product is mentioned, and position shows its average place in the carousel when it appears (lower is better). Peec AI also shows the merchants listed for each product, the competing products next to yours, and the shopping queries ChatGPT runs to build the carousel.
Start a free trial or book a demo of Peec AI to learn more.
